Prediction of Remaining Service Life of Bridge Decks Using Machine Learning

This paper demonstrates that it is feasible to use either the k-nearest-neighbor instance-based learning (IBL) technique or the inductive learning (IL) technique for engineering applications in classification and prediction problems such as estimating the remaining service life of bridge decks. It is shown that IBL is more efficient than IL: The best achieved percentages of correctly classified instances are 50% as generated by k-nearest-neighbor IBL and 41.8% when generated by the C4.5/IL learning algorithm. From a machine learning (ML) standpoint both these values are considered low, but this is attributed to the fact that the deterioration model used to compute the remaining service life turned out to be inadequate. It is based on a methodology developed under the Strategic Highway Research Program (SHRP) for life-cost analysis of concrete bridges relative to reinforcement corrosion. Actual bridge deck surveys were obtained from the Kansas Department of Transportation that include the type of attributes needed for the SHRP methodology. The experimentation with the ML algorithms reported here also describes the experience one may go through when faced with an imperfect model, or with incomplete data or missing attributes.